ACCORD-Net: anatomically consistent cross-modality diffusion learning for low-dose brain image reconstruction

Proper delineation of brain tumors and lesion boundaries is essential for early diagnosis, surgical planning and monitoring of treatment. Low-dose computed tomography (CT) and under-sampled magnetic resonance imaging (MRI) often suffer from noise, low contrast, and structural distortions, which can compromise diagnostic accuracy. To tackle these problems, this study presents a regularized diffusion probabilistic reconstruction framework (ReDiffuse-Net) that integrates cross-modality learning to improve low-dose brain imaging. It combines a denoising diffusion probabilistic model with data-consistency and structural-consistency constraints, using anatomical priors from MRI to aid the reconstruction of low-dose CT images. The forward diffusion process simulates real degradation patterns, and the reverse denoising process gradually recovers the fine-grained and structurally accurate brain structure in a multimodal supervised manner. Results from experiments performed on the CT of the Brain – Object Detection dataset and the UCSF-PDGM MRI dataset show significant improvement in the reconstruction quality achieved by ReDiffuse-Net, with up to 18.6% higher PSNR and 14.3% higher SSIM compared to the state-of-the-art CNN–, GAN-, and Transformer-based methods. Reconstructed images show better tumor boundaries, more prominent gray–white matter differences, and better image interpretability, validating the potential of ReDiffuse-Net as a powerful, multimodal reconstruction method for brain tumor imaging, radiotherapy planning, and disease progression evaluation.